@grida/ai-models 0.0.0-stage → 1.20261011.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/LICENSE ADDED
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 Grida
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
package/PROVENANCE.md ADDED
@@ -0,0 +1,681 @@
1
+ # Catalogue research notes
2
+
3
+ Source-specific qualifications and references for the values in
4
+ [`models.json`](./models.json). The JSON holds the values; these notes explain
5
+ where they came from and any caveat a value cannot carry itself. Update a note
6
+ in the same change as the fact it qualifies.
7
+
8
+ Each `##` heading names the entry or declaration a note qualifies. The headings
9
+ are code fragments carried over from the TypeScript literals the catalogue was
10
+ first written as in [gridaco/grida](https://github.com/gridaco/grida), so they
11
+ keep that historical spelling (for example `imageInputMimes` rather than
12
+ `image_input_mimes`).
13
+
14
+ ## `const OPENAI_IMAGE_INPUT_MIMES = [`
15
+
16
+ Provider-family capabilities are kept private so catalogue entries remain
17
+ explicit while sharing one source-backed value. Do not derive these from
18
+ `multimodal`; a future model may be multimodal without a documented image
19
+ input format set.
20
+ https://developers.openai.com/api/docs/guides/images-vision#image-input-requirements
21
+
22
+ ## `const ANTHROPIC_IMAGE_INPUT_MIMES = [`
23
+
24
+ https://platform.claude.com/docs/en/build-with-claude/vision#supported-formats
25
+
26
+ ## `const GOOGLE_IMAGE_INPUT_MIMES = [`
27
+
28
+ https://ai.google.dev/gemini-api/docs/image-understanding#supported-image-formats
29
+
30
+ ## `const OPENAI_LONG_CONTEXT_PRICING = {`
31
+
32
+ OpenAI bills the full request at these multipliers once its total input
33
+ exceeds 272K tokens. The same rule is published for GPT-5.5, every
34
+ GPT-5.6 family member, and GPT-6 Astra.
35
+ https://developers.openai.com/api/docs/models/gpt-5.5
36
+ https://developers.openai.com/api/docs/models/gpt-5.6-sol
37
+ https://developers.openai.com/api/docs/models/gpt-6-astra
38
+
39
+ ## `"openai/gpt-5.6-sol": {`
40
+
41
+ Base rates; OPENAI_LONG_CONTEXT_PRICING represents the request-wide
42
+ band that applies above 272K total input tokens.
43
+
44
+ Sol is cheaper than GPT-5.5, the card directly above it. It was
45
+ introduced carrying 5.5's rates verbatim; these are OpenAI's own.
46
+ https://developers.openai.com/api/docs/models/gpt-5.6-sol
47
+
48
+ ## `"openai/gpt-6-astra": {`
49
+
50
+ OpenAI's September 3 introduction remained a limited rollout. The
51
+ September 4 date below is the exact Vercel AI Gateway route's broad availability,
52
+ which is the release fact relevant to this route card.
53
+ https://openai.com/products/release-notes/
54
+ https://vercel.com/ai-gateway/models/gpt-6-astra
55
+
56
+ ## `"anthropic/claude-opus-5.5": {`
57
+
58
+ Anthropic API ID `claude-opus-5-5`; this is the Vercel AI Gateway ID, whose
59
+ model feed lists the same rates. Cache reads are 0.05x base input ($0.20);
60
+ `cache_write` is the 5-minute write ($5; the 1-hour write is $8). Released
61
+ 2026-09-22 as generally available; only its fast mode is a research preview.
62
+ https://platform.claude.com/docs/en/models/opus-5-5/overview
63
+ https://platform.claude.com/docs/en/about-claude/pricing
64
+ https://vercel.com/ai-gateway/models/claude-opus-5.5
65
+
66
+ Claude Opus 5 stays catalogued. Opus 5.5 is cheaper, but it is not a drop-in
67
+ successor: thinking can't be disabled, forced tool use (`tool_choice`
68
+ `any`/`tool`) returns an error, and `computer_20251124` is not accepted on
69
+ the Claude API and Google Cloud. Opus 5 is still the choice for those.
70
+ https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5
71
+
72
+ ## `"anthropic/claude-haiku-5.5": {`
73
+
74
+ Anthropic API ID `claude-haiku-5-5`; this is the Vercel AI Gateway ID. Priced
75
+ by prompt length: a request whose prompt is over 100,000 tokens, counting all
76
+ input tokens including cache reads and writes, pays $0.50 input, $2.50
77
+ output, $0.05 cache read and $0.625 5-minute cache write, 5x each base rate.
78
+ `long_context` records that band; its input multiplier also applies to cache
79
+ reads and writes. The gateway's model feed publishes the same tiers. The
80
+ 1.1x US-only inference multiplier (`inference_geo`) also applies to the
81
+ higher band.
82
+ https://platform.claude.com/docs/en/models/haiku-5-5/overview
83
+ https://platform.claude.com/docs/en/about-claude/pricing
84
+ https://vercel.com/ai-gateway/models/claude-haiku-5.5
85
+
86
+ ## `"anthropic/claude-sonnet-5": {`
87
+
88
+ $2/$10 is the standard rate, not a live discount: it launched as an
89
+ introductory rate and Anthropic made it permanent, cancelling the
90
+ announced step up to $3/$15. Do not restore the higher card.
91
+ https://platform.claude.com/docs/en/about-claude/pricing
92
+
93
+ ## `"anthropic/claude-fable-5.1": {`
94
+
95
+ Same input/output/cacheWrite card as Claude Fable 5, with cache reads
96
+ cut to $0.25/MTok. Not a drop-in successor — forced tool choice
97
+ (`tool_choice` `any`/`tool`) is rejected here — which is why Fable 5
98
+ stays catalogued rather than being removed.
99
+ https://platform.claude.com/docs/en/models/fable-5-1/overview
100
+
101
+ ## `"anthropic/claude-fable-5": {`
102
+
103
+ Unlike Fable 5.1, this version accepts forced tool choice.
104
+
105
+ ## `"anthropic/claude-opus-5": {`
106
+
107
+ Drop-in successor to Opus 4.8 at the same rate card.
108
+
109
+ ## `"google/gemini-3.8-flash": {`
110
+
111
+ Google's cache model is read + hourly storage (no one-time write
112
+ premium that matches `cacheWrite` semantics), so the field is omitted.
113
+
114
+ Steady-state rates. Google is promoting $0.75 in / $3.75 out / $0.075
115
+ cacheRead through 2026-12-31; these are the prices that apply from
116
+ 2027-01-01. Checking Google's page before then will show the lower
117
+ set — that is the promotion, not a correction. Gemini 3.8 Flash is GA,
118
+ and this exact id is live on both Vercel AI Gateway and OpenRouter.
119
+ https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
120
+ https://ai.google.dev/gemini-api/docs/pricing
121
+ https://vercel.com/ai-gateway/models/gemini-3.8-flash
122
+ https://openrouter.ai/google/gemini-3.8-flash
123
+
124
+ ## `"google/gemini-3.7-flash": {`
125
+
126
+ 3.8 improves accuracy and reliability at the same rate, but Google
127
+ still recommends 3.7 when compute efficiency matters because 3.8 can
128
+ consume more tokens.
129
+ https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-3-8-flash
130
+
131
+ ## `longContext: {`
132
+
133
+ Google bills the full request at $4 in / $18 out / $0.40
134
+ cacheRead above 200K tokens. `inputMultiplier` covers every
135
+ input bucket, so cacheRead 0.2 x 2 lands on $0.40 without a
136
+ separate field.
137
+ https://ai.google.dev/gemini-api/docs/pricing
138
+
139
+ ## `const GPT_IMAGE_2_5_FAL_PRICING: PerTokenPricing = {`
140
+
141
+ Both 2.5 variants use this fal meter. Keep the separate text-output
142
+ rate: fal publishes it even though its result schema exposes only images.
143
+ https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image
144
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image
145
+
146
+ ## `const GPT_IMAGE_2_5_VERCEL_AI_GATEWAY_PRICING: PerTokenPricing = {`
147
+
148
+ Each serving provider publishes its own meter; do not copy fal's extra
149
+ image-cache/text-output rates into providers that do not advertise them.
150
+ https://ai-gateway.vercel.sh/v1/models (verified 2026-09-09 KST)
151
+
152
+ ## `const GPT_IMAGE_2_5_OPENROUTER_PRICING: PerTokenPricing = {`
153
+
154
+ Both variants' /api/v1/images/models/{id}/endpoints publish this meter.
155
+ https://openrouter.ai/api/v1/images/models/openai/gpt-image-2.5-flare/endpoints
156
+ https://openrouter.ai/api/v1/images/models/openai/gpt-image-2.5-sunburst/endpoints
157
+
158
+ ## `"openai/gpt-image-2": {`
159
+
160
+ ---
161
+
162
+ OpenAI
163
+ -----------------------------------------------------------------
164
+
165
+ https://developers.openai.com/api/docs/models/gpt-image-2
166
+
167
+ ## `transparent_background: true,`
168
+
169
+ Native transparency added 2026-08-20; provider exposure differs.
170
+ https://developers.openai.com/api/docs/changelog
171
+
172
+ ## `providers: {`
173
+
174
+ ids/prices verified 2026-06-29, see github.com/gridaco/grida/issues/908
175
+
176
+ ## `transparent_background: null,`
177
+
178
+ This provider's page does not establish background support.
179
+ https://vercel.com/ai-gateway/models/gpt-image-2 (2026-09-09)
180
+
181
+ ## `transparent_background: false,`
182
+
183
+ Published background enum is auto | opaque (2026-09-09).
184
+ https://openrouter.ai/api/v1/images/models/openai/gpt-image-2/endpoints
185
+
186
+ ## `references: { id: "openai/gpt-image-2", max: 16 },`
187
+
188
+ input_references advertised by OpenRouter (0–16), 2026-07-01.
189
+
190
+ ## `},`
191
+
192
+ Inherits native transparency: fal-ai/gpt-image-2's published
193
+ background enum includes transparent (verified 2026-09-09).
194
+
195
+ ## `"openai/gpt-image-2.5-flare": {`
196
+
197
+ Released as two distinct models. All three providers now list both
198
+ variants; fal separates generation and edit endpoint ids, while
199
+ OpenRouter advertises input_references on the same id (2026-09-09 KST).
200
+
201
+ ## `transparent_background: true,`
202
+
203
+ Both fal generation and edit schemas expose background=transparent.
204
+ https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image/api
205
+ https://fal.ai/models/openai/gpt-image-2.5/flare/edit/api
206
+
207
+ ## `url: "https://vercel.com/ai-gateway/models/gpt-image-2.5-flare",`
208
+
209
+ The model page exposes background=transparent; reference input
210
+ is not established by the Vercel AI Gateway feed (2026-09-09 KST).
211
+
212
+ ## `transparent_background: false,`
213
+
214
+ Endpoint schema's background enum is auto | opaque.
215
+
216
+ ## `constraints: {`
217
+
218
+ Model-specific fal documentation overrides its generic ImageSize
219
+ component's wider bounds. The API also accepts auto-sized output.
220
+ https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image/api
221
+
222
+ ## `avg_cost_usd: 0.055,`
223
+
224
+ High 1024² output estimate is $0.05268, plus a small input allowance.
225
+ Not a fixed per-image price: actual cost depends on all billed tokens.
226
+ https://developers.openai.com/api/docs/guides/image-generation
227
+
228
+ ## `transparent_background: true,`
229
+
230
+ Both fal generation and edit schemas expose background=transparent.
231
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image/api
232
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/edit/api
233
+
234
+ ## `url: "https://vercel.com/ai-gateway/models/gpt-image-2.5-sunburst",`
235
+
236
+ The model page exposes background=transparent; reference input
237
+ is not established by the Vercel AI Gateway feed (2026-09-09 KST).
238
+
239
+ ## `transparent_background: false,`
240
+
241
+ Endpoint schema's background enum is auto | opaque.
242
+
243
+ ## `constraints: {`
244
+
245
+ https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image/api
246
+
247
+ ## `"openai/gpt-image-1.5": {`
248
+
249
+ https://developers.openai.com/api/docs/models/gpt-image-1.5
250
+
251
+ ## `constraints: null,`
252
+
253
+ Preset-only — provider rejects arbitrary sizes.
254
+
255
+ ## `pricing: {`
256
+
257
+ https://developers.openai.com/api/docs/models/gpt-image-1.5
258
+
259
+ ## `"openai/gpt-image-1-mini": {`
260
+
261
+ https://developers.openai.com/api/docs/models/gpt-image-1-mini
262
+
263
+ ## `constraints: null,`
264
+
265
+ Preset-only — provider rejects arbitrary sizes.
266
+
267
+ ## `pricing: {`
268
+
269
+ https://developers.openai.com/api/docs/models/gpt-image-1-mini
270
+
271
+ ## `"google/gemini-3.1-flash-image-preview": {`
272
+
273
+ ---
274
+
275
+ Google (multimodal LLMs with native image output)
276
+ -----------------------------------------------------------------
277
+
278
+ python .tools/model_info.py --image gemini-3.1-flash-image
279
+ Vercel AI Gateway pricing: $0.50/MTok input, $3.00/MTok output
280
+
281
+ ## `providers: {`
282
+
283
+ "Nano Banana 2"; ids/prices verified 2026-06-29, see issues/908
284
+
285
+ ## `vercel: {`
286
+
287
+ Vercel AI Gateway serves the graduated `google/gemini-3.1-flash-image`
288
+ and the `-preview` alias at identical rates (feed, 2026-09-02).
289
+ Bindings call the graduated id; the canonical key above stays
290
+ `-preview` because it is persisted in selections and published
291
+ in the catalogue snapshot — renaming it is a separate change.
292
+
293
+ ## `references: { id: "google/gemini-3.1-flash-image", max: 14 },`
294
+
295
+ input_references advertised by OpenRouter (0–14), 2026-07-01.
296
+
297
+ ## `fal: {`
298
+
299
+ fal's graduated endpoint is `fal-ai/nano-banana-2`; same $0.08
300
+ per 1K image as the `-preview` endpoint (2K ×1.5, 4K ×2).
301
+
302
+ ## `"google/gemini-3-pro-image": {`
303
+
304
+ python .tools/model_info.py --image gemini-3-pro-image
305
+ Vercel AI Gateway pricing: $2.00/MTok input, $12.00/MTok output
306
+
307
+ ## `providers: {`
308
+
309
+ "Nano Banana Pro"; ids/prices verified 2026-06-29, see issues/908
310
+
311
+ ## `references: { id: "google/gemini-3-pro-image-preview", max: 14 },`
312
+
313
+ input_references advertised by OpenRouter (0–14), 2026-07-01.
314
+
315
+ ## `"google/gemini-3.1-flash-lite-image": {`
316
+
317
+ "Nano Banana 2 Lite" — GA 2026-06-30. The cost/speed tier of the 3.1
318
+ Flash family: ~half of Nano Banana 2's meter, and 1K-only output
319
+ (2K/4K unsupported — the differentiator). Vercel AI Gateway + OpenRouter both
320
+ meter it at $0.25/$1.50 (verified 2026-07-01); fal id not verified,
321
+ so left out. OpenRouter doesn't advertise input_references for the
322
+ Lite (t2i only per its model page), so no `references` (TOOL-DESIGN:
323
+ no unverified capability).
324
+
325
+ ## `avg_cost_usd: 0.034,`
326
+
327
+ Published 1K per-image cost (the card default): $0.034 = 1120 tokens
328
+ × $30/1M image-output (Google/Vercel AI Gateway changelog, 2026-07-01). The
329
+ budget meter charges this per image, so it must be the real cost.
330
+
331
+ ## `"bfl/flux-2-pro": {`
332
+
333
+ ---
334
+
335
+ Black Forest Labs (via Vercel AI Gateway)
336
+ -----------------------------------------------------------------
337
+
338
+ https://vercel.com/docs/ai-gateway/capabilities/image-generation/ai-sdk
339
+ https://docs.bfl.ml/pricing
340
+
341
+ ## `providers: {`
342
+
343
+ All three providers meter $0.03 per megapixel (Vercel AI Gateway model page,
344
+ OpenRouter endpoint `cost_usd`/megapixel, fal "first megapixel");
345
+ represented as flat at the 1MP baseline. The Vercel AI Gateway binding shipped
346
+ at $0.06 — a 2x hosted over-billing — corrected 2026-09-02. The
347
+ Vercel AI Gateway feed carries no `pricing` for BFL cards, so the page is the
348
+ source.
349
+
350
+ ## `references: { id: "black-forest-labs/flux.2-pro", max: 8 },`
351
+
352
+ input_references advertised by OpenRouter (0–8), 2026-07-01.
353
+
354
+ ## `"bfl/flux-2-max": {`
355
+
356
+ ---
357
+
358
+ Black Forest Labs — Flux 2 Max
359
+ -----------------------------------------------------------------
360
+
361
+ BFL's top Flux 2 line (2025-12-16). $0.07 per megapixel on all three
362
+ (Vercel AI Gateway model page — the feed carries no BFL pricing; OpenRouter
363
+ `cost_usd`/megapixel; fal "first megapixel", +$0.03 each additional).
364
+ Represented as flat at the 1MP baseline. Verified 2026-09-02.
365
+
366
+ ## `references: { id: "black-forest-labs/flux.2-max", max: 8 },`
367
+
368
+ OpenRouter `supported_parameters.input_references` 0–8
369
+ (2026-09-02). As on Flux 2 Pro.
370
+
371
+ ## `constraints: { min_edge: 256, max_edge: 1440 },`
372
+
373
+ Same envelope as Flux 2 Pro pending a vendor spec for Max.
374
+
375
+ ## `providers: {`
376
+
377
+ $0.04 on both providers: Vercel AI Gateway feed `pricing.image` + fal's model
378
+ page ("Fixed $0.04 cost per image edit"), verified 2026-09-02.
379
+
380
+ ## `"bytedance/seedream-5.0-pro": {`
381
+
382
+ ---
383
+
384
+ ByteDance — Seedream 5.0 Pro
385
+ -----------------------------------------------------------------
386
+
387
+ Vercel AI Gateway feed `pricing.image` $0.035 (the model page's rate table shows
388
+ $0.04 while its copy says $0.035 — the feed is the billing contract);
389
+ OpenRouter `cost_usd` $0.045 at 1K ($0.09 high-res, +$0.003 per
390
+ input image, not modelled); fal $0.0675 for ≤1536² area, $0.135 up
391
+ to 2048². Represented at the 1024² baseline. Verified 2026-09-02.
392
+
393
+ ## `references: { id: "bytedance-seed/seedream-5-0-pro", max: 14 },`
394
+
395
+ OpenRouter `supported_parameters.input_references` 0–14
396
+ (2026-09-02). Same endpoint as t2i, as on 4.5.
397
+
398
+ ## `constraints: {`
399
+
400
+ fal: total pixels between 1024x1024 and 2048x2048, aspect ratio
401
+ between 1:16 and 16:1 — an area bound, not a per-edge one, so a
402
+ 512x2048 request is in-envelope and a 2048x2048 one is the ceiling.
403
+
404
+ ## `"bytedance/seedream-5.0-lite": {`
405
+
406
+ ---
407
+
408
+ ByteDance — Seedream 5.0 Lite
409
+ -----------------------------------------------------------------
410
+
411
+ Universal: $0.035/img on all three (Vercel AI Gateway feed `pricing.image`,
412
+ OpenRouter `cost_usd`, fal page payload), verified 2026-09-02. A
413
+ 2K–4K model: fal scales requests below 2560x1440 up to its floor.
414
+
415
+ ## `references: { id: "bytedance-seed/seedream-5-0-lite", max: 14 },`
416
+
417
+ OpenRouter `supported_parameters.input_references` 0–14
418
+ (2026-09-02).
419
+
420
+ ## `constraints: { min_pixels: 3_686_400, max_pixels: 16_777_216 },`
421
+
422
+ fal: total pixels between 2560x1440 and 4096x4096 (requests below
423
+ the floor are scaled up to it); Vercel AI Gateway and OpenRouter serve 2K/4K
424
+ only. An area envelope, so the default is a 2K request.
425
+
426
+ ## `"bytedance/seedream-4.5": {`
427
+
428
+ ---
429
+
430
+ ByteDance — Seedream 4.5
431
+ -----------------------------------------------------------------
432
+
433
+ $0.04/img on all three providers (re-verified 2026-09-02).
434
+
435
+ ## `references: { id: "bytedance-seed/seedream-4.5", max: 14 },`
436
+
437
+ i2i verified live 2026-07-01 (OpenRouter /api/v1/images
438
+ input_references; same endpoint as t2i). max from
439
+ supported_parameters.input_references.
440
+
441
+ ## `"xai/grok-imagine-image-2.0": {`
442
+
443
+ ---
444
+
445
+ SpaceXAI — Grok Imagine Image 2.0
446
+ -----------------------------------------------------------------
447
+
448
+ Tiered by quality (low/medium) × resolution (1K/2K); the same four
449
+ rates on all three providers (Vercel AI Gateway feed
450
+ `image_dimension_quality_pricing`, OpenRouter `cost_usd` variants,
451
+ fal page). OpenRouter also bills $0.01 per input image (not
452
+ modelled). Verified 2026-09-02.
453
+
454
+ ## `references: { id: "x-ai/grok-imagine-image-2.0", max: 3 },`
455
+
456
+ OpenRouter `supported_parameters.input_references` 0–3
457
+ (2026-09-02). OR bills $0.01 per input image on top (not modelled).
458
+
459
+ ## `"meta/muse-image-1.0": {`
460
+
461
+ ---
462
+
463
+ Meta — Muse Image 1.0
464
+ -----------------------------------------------------------------
465
+
466
+ Meta's agentic image model (2026-08-26): $0.01/img on Vercel AI Gateway (feed
467
+ `pricing.image`) and fal (page payload). OpenRouter lists it but
468
+ exposes no serving endpoint.
469
+ fal exposes aspect ratio only (no size control). Verified 2026-09-02.
470
+
471
+ ## `constraints: null,`
472
+
473
+ Muse picks output dimensions from the aspect ratio; neither
474
+ provider exposes a size envelope, so none is claimed.
475
+
476
+ ## `"recraft/recraft-v4.1": {`
477
+
478
+ ---
479
+
480
+ Recraft — V4.1
481
+ -----------------------------------------------------------------
482
+
483
+ Universal: $0.035/img raster on every provider (Vercel AI Gateway feed
484
+ `pricing.image`, OpenRouter endpoint `cost_usd`, fal page payload),
485
+ verified 2026-09-02. Vector styles are $0.08 and a separate route on
486
+ fal/OpenRouter (`.../text-to-vector`, `recraft-v4.1-vector`) — not
487
+ catalogued; `styles: null` here means the raster route only.
488
+ OpenRouter org slug is `recraft`, not `recraft-ai`.
489
+
490
+ ## `references: { id: "recraft/recraft-v4.1", max: 1 },`
491
+
492
+ OpenRouter `supported_parameters.input_references` 0–1
493
+ (2026-09-02).
494
+
495
+ ## `"recraft/recraft-v3": {`
496
+
497
+ ---
498
+
499
+ Recraft — V3
500
+ -----------------------------------------------------------------
501
+
502
+ $0.04/img raster on every provider (re-verified 2026-09-02; Recraft's
503
+ own pricing table agrees).
504
+
505
+ ## `pricing: { type: "per_run_flat", usd: 0.04 },`
506
+
507
+ Source: replicate.com/google/lyria-3 — "$0.04 per output audio file"
508
+
509
+ ## `pricing: { type: "per_run_flat", usd: 0.08 },`
510
+
511
+ Source: replicate.com/google/lyria-3-pro — "$0.08 per output audio file"
512
+
513
+ ## `pricing: {`
514
+
515
+ API: 100 credits when duration is automatic, or 11 credits/s when set.
516
+
517
+ ## `pricing: { type: "per_1000_characters", usd: 0.1 },`
518
+
519
+ Source: elevenlabs.io/pricing/api — $0.10 per 1,000 characters.
520
+
521
+ ## `"tripo/rig-v2.5": {`
522
+
523
+ Follow the documented compatibility table, not its contradictory biped example.
524
+
525
+ ## `"google/veo-3.1": {`
526
+
527
+ ---
528
+
529
+ Google — Veo 3.1
530
+ -----------------------------------------------------------------
531
+
532
+ ## `vercel: {`
533
+
534
+ Vercel AI Gateway — vercelAiGateway.videoModel(id), image-to-video. The
535
+ Vercel AI Gateway meters both audio modes and sells 4K; the matrix is
536
+ identical to fal's. Verified against Vercel AI Gateway's own
537
+ /v1/models feed (`video_duration_pricing`) on 2026-09-02 — the
538
+ previous card claimed "audio-on only, ≤1080p", which the feed
539
+ contradicts.
540
+
541
+ `silent` is catalogued as provider truth but is not yet
542
+ reachable on Grida's hosted route: the video wire carries no
543
+ audio-mode field, so `generateVideo` prices the card's default
544
+ mode ([`editor/lib/ai/server.ts`](https://github.com/gridaco/grida/blob/main/editor/lib/ai/server.ts) in gridaco/grida). `"4k"` IS reachable — the
545
+ request path maps a 2160 short edge to that label and today
546
+ rejects it for want of a rate.
547
+ https://vercel.com/ai-gateway/models/veo-3.1-generate-001
548
+
549
+ ## `fal: {`
550
+
551
+ fal.ai — image-to-video endpoint (capability is keyed into the id;
552
+ t2v is a separate `fal-ai/veo3.1` endpoint, not catalogued). Rate
553
+ matrix is identical to the Vercel AI Gateway's: $0.40/s audio and
554
+ $0.20/s silent at 720p/1080p, $0.60/$0.40 at 4K.
555
+ https://fal.ai/models/fal-ai/veo3.1/image-to-video
556
+
557
+ ## `openrouter: {`
558
+
559
+ OpenRouter — async `/api/v1/videos` (job → poll → unsigned url).
560
+ Text-to-video + image-to-video; native audio. "from $0.40/s"
561
+ (verified 2026-06-29, https://openrouter.ai/google/veo-3.1).
562
+
563
+ ## `"google/veo-3.1-fast": {`
564
+
565
+ ---
566
+
567
+ Google — Veo 3.1 Fast
568
+ -----------------------------------------------------------------
569
+
570
+ Same envelope as Veo 3.1 (16:9/9:16, 4/6/8s, native audio, ≤4K) at
571
+ ~2.7x lower cost. Rate matrix is identical on Vercel AI Gateway and fal —
572
+ Vercel AI Gateway feed `video_duration_pricing` and fal's stated per-second
573
+ rates, verified 2026-09-02.
574
+
575
+ ## `"google/veo-3.1-lite": {`
576
+
577
+ ---
578
+
579
+ Google — Veo 3.1 Lite
580
+ -----------------------------------------------------------------
581
+
582
+ The budget Veo: 720p/1080p only, 4/6/8s, native audio. Rates
583
+ identical on Vercel AI Gateway and fal (verified 2026-09-02). Vercel AI Gateway
584
+ supports both text and image input (FAQ verified 2026-09-07); the fal binding
585
+ below requires an image. Hosted GG's narrower wire is a host concern.
586
+
587
+ ## `"alibaba/wan-3.0": {`
588
+
589
+ ---
590
+
591
+ Alibaba — Wan 3.0
592
+ -----------------------------------------------------------------
593
+
594
+ Per-second by resolution, audio bundled into the rate (Vercel AI Gateway
595
+ feed has no audio axis; fal exposes an `audio` toggle but bills the
596
+ same). Identical on Vercel AI Gateway and fal, verified 2026-09-02. 2–30s.
597
+
598
+ ## `"bytedance/seedance-2.0": {`
599
+
600
+ ByteDance — Seedance 2.0
601
+ -----------------------------------------------------------------
602
+
603
+ ## `providers: {`
604
+
605
+ NO Vercel AI Gateway binding, deliberately. Vercel AI Gateway serves
606
+ `bytedance/seedance-2.0` but meters it PER TOKEN
607
+ (`video_token_pricing`: $7.00/MTok at 480p/720p, $7.70 at 1080p,
608
+ $4.00 at 4K; reduced with video input; "minimum token floors based
609
+ on output duration"). `PerSecondPricing` cannot express that, the
610
+ hosted route pre-prices rate×duration, and ByteDance publishes no
611
+ tokens-per-second figure — so there is no honest per-second rate,
612
+ and the per-second Vercel AI Gateway binding this card shipped with was
613
+ invented. Withheld until the hosted path can meter post-flight;
614
+ hosted requests fail with "not available on the hosted provider".
615
+ See gridaco/grida#1019 (Blocker A). Feed checked 2026-09-02.
616
+
617
+ ## `fal: {`
618
+
619
+ fal — image-to-video. fal also meters tokens underneath
620
+ ($0.014/1K), but states a per-second price for the two
621
+ resolutions it prices: $0.3034/s @720p, $0.682/s @1080p, audio
622
+ bundled. Those are fal's own figures, not a conversion of ours.
623
+ https://fal.ai/models/bytedance/seedance-2.0/image-to-video
624
+
625
+ ## `openrouter: {`
626
+
627
+ OpenRouter — async `/api/v1/videos`. Flat $0.06726/s (verified
628
+ 2026-06-29, https://openrouter.ai/bytedance/seedance-2.0) — far
629
+ below Vercel AI Gateway's per-resolution rate (the proprietary-pricing-
630
+ diverges finding, #908). No separate silent meter surfaced.
631
+
632
+ ## `"bytedance/seedance-2.5": {`
633
+
634
+ ---
635
+
636
+ ByteDance — Seedance 2.5
637
+ -----------------------------------------------------------------
638
+
639
+ Newer generation (2026-07-31), NOT a drop-in successor: ~55% more
640
+ per token on Vercel AI Gateway ($10.70/MTok at 480p/720p, $11.70 at 1080p
641
+ vs 2.0's $7.00/$7.70), ~56–71% more per second on fal, and no 4K.
642
+ It buys 4–30s clips, video-editing and extend-video. 2.0 is cheaper
643
+ and serves 4K. Vercel AI Gateway bills tokens, so there is no comparable
644
+ per-second Vercel AI Gateway meter here. fal states per-second prices, audio bundled.
645
+
646
+ ## `fal: {`
647
+
648
+ fal's stated rates: $0.2205/s @480p, $0.4730/s @720p, $1.164/s
649
+ @1080p (fal meters $0.0214/1K tokens underneath). 2026-09-02.
650
+ https://fal.ai/models/bytedance/seedance-2.5/image-to-video
651
+
652
+ ## `"xai/grok-imagine-video-1.5": {`
653
+
654
+ ---
655
+
656
+ SpaceXAI — Grok Imagine Video 1.5
657
+ -----------------------------------------------------------------
658
+
659
+ Image-to-video only (no t2v, per SpaceXAI docs); native lip-synced audio
660
+ bundled into the rate. Per-second by resolution, identical on Vercel AI Gateway
661
+ (no markup) and fal: $0.08/s @480p, $0.14/s @720p, $0.25/s @1080p.
662
+ Both also bill $0.01 per input image, captured separately from the
663
+ output meter.
664
+
665
+ ## `vercel: {`
666
+
667
+ Vercel AI Gateway — image-to-video; mirrors SpaceXAI's list price (no markup).
668
+ Vercel AI Gateway namespaces every SpaceXAI model under `spacexai/`, so the
669
+ call id deliberately differs from this card's canonical `xai/` id.
670
+ https://vercel.com/changelog/grok-imagine-video-1-5-on-ai-gateway
671
+
672
+ ## `fal: {`
673
+
674
+ fal.ai — image-to-video endpoint; same per-second rate.
675
+ https://fal.ai/models/xai/grok-imagine-video/v1.5/image-to-video
676
+
677
+ ## `pricing: { type: "per_token_input", input: 0.15 },`
678
+
679
+ VERIFY before prod: models.dev lists no per-1M price for
680
+ gemini-embedding-2 yet; using the gemini-embedding-001 family
681
+ input rate ($0.15 / 1M) as a conservative stand-in.